Le comportement IA que personne n'avait prédit (actus IA)

This week in the AI ​​Review: a deep dive into Ring-Zero, the paper that demonstrates how zero reinforcement learning can uncover reasoning behaviors at the scale of a trillion parameters. We discuss DeepSeek-R1-Zero, Rich Sutton's "bitter lesson," context anxiety, parallel reasoning, self-verification, and what it really reveals about modern artificial intelligence. Also on the agenda: Kimi K3 and Qwen 3.8 Max, two Chinese open-weight models that are catching up to the frontier; Inkling, Thinking Machines' first large-scale American open-weight model; PrismML Bonsai 27B, a compressed model designed to run on iPhones; and three rapid innovations in AI video, real-time avatars, and computer vision. 🎓 AI and Second Brain Masterclass: https://prisme.one/inscription-master... 00:00 Ring-Zero, the AI ​​that learns to think on its own 02:49 DeepSeek-R1-Zero and the zero RL 04:18 The transition to 1 trillion parameters 05:07 Emergent behaviors of the model 06:18 Context anxiety 08:39 The Bitter Lesson: scale beats human engineering 10:54 News 1: Kimi K3, the Chinese model catching up to Claude Fable 11:46 News 2: Qwen 3.8 Max, Alibaba accelerates 12:41 Why open weights are changing AI sovereignty 13:05 News 3: Inkling, the first major American open weights firm 14:55 News 4: PrismML, Apple is preparing local AI on iPhone 16:20 Masterclass Free AI 4:37 PM News Item 5: Live Avatar, the real-time AI avatar 5:08 PM News Item 6: Wan-Dancer, 3 minutes of generated dance 5:45 PM News Item 7: GenCeption, Google DeepMind rethinks computer vision 🗞️ SOURCES: Deep Dive - Ring-Zero, zero RL and the emergence of reasoning Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning https://arxiv.org/html/2607.12395 The Bitter Lesson, Rich Sutton http://www.incompleteideas.net/IncIde... News Item 1 - Kimi K3 and Qwen 3.8 Max, the Chinese open weights catch up with the frontier Announces Kimi K3, Moonshot A.I. https://www.kimi.com/blog/kimi-k3 Analysis Kimi K3, Simon Willison https://simonwillison.net/2026/Jul/16... Benchmarks Kimi K3, theAIsearch    • Kimi K3, dancing waifus, robot UFC, song t...   Qwen 3.8 Max Preview, The Decoder https://the-decoder.com/alibabas-qwen... News 2 - Thinking Machines Inkling, the first large American open weights Official announcement Inkling, Thinking Machines https://thinkingmachines.ai/news/intr... Model card Inkling https://thinkingmachines.ai/model-car... Inkling Presentation, Hugging Face https://huggingface.co/blog/thinkingm... Inkling Analysis, Simon Willison https://simonwillison.net/2026/Jul/16... News 3 - PrismML Bonsai 27B, a 27B model that runs on iPhone Bonsai 27B Announcement, PrismML https://prismml.com/news/bonsai-27b 9to5Mac Article on Bonsai 27B https://9to5mac.com/2026/07/14/prismm... Apple Acquisition Discussion AppleInsider https://appleinsider.com/articles/26/... Express Radar - Live Avatar Paper Live Avatar https://arxiv.org/abs/2512.04677 GitHub LiveAvatar https://github.com/Alibaba-Quark/Live... Radar express - Wan-Dancer Wan-Dancer project page https://humanaigc.github.io/wan-dance... Express radar - GenCeption Paper GenCeption, Google DeepMind https://arxiv.org/html/2607.09024v1